ST-FlowNet: A lightweight framework for long-term spatio-temporal flow field prediction.

Abstract

Unsteady flows are ubiquitous in critical fields such as aerospace and energy. Due to their high-dimensional and complex properties that evolve across space and time, spatio-temporal flow field prediction remains a challenge in exploring the underlying physical mechanisms. Traditional numerical methods tend to incur high computational costs due to iterative calculations. Existing intelligent approaches offer efficient alternatives but suffer from error accumulation during the long-term prediction process, leading to a sharp decline in accuracy over time. In this paper, we propose ST-FlowNet, a lightweight framework for accurate long-term spatio-temporal flow field prediction. Specifically, we introduce proper orthogonal decomposition (POD) to decouple flow characteristics, thus reducing modeling complexity and feature extraction costs. To alleviate error accumulation, we develop an attention-enhanced model guided by a well-designed physics-informed loss function for long-term temporal prediction. Experimental results demonstrate that ST-FlowNet accurately simulates the long-term evolution of unsteady flows and obtains a two-order-of-magnitude speedup compared to the traditional numerical simulation. It achieves state-of-the-art prediction accuracy and generalization capability while maintaining a minimal parameter size compared to existing intelligent models.